Over on the 10x10 computer draughts forum, there were tons of replies to Fabien's introductory post about Scan patterns:fierz wrote: ↑Tue Mar 23, 2021 12:28 pm Thanks for your answers!
I should maybe mention first that I never found any good accessible explanations of how to do a pattern optimization (just as I have never found any good accessible explanation of new concepts such as NNUE); until a few weeks ago I was using a Texel-kind of tuning method for first 400 then about 4000 parameters (3-backrank instead of 2-backrank = 4096-256 additional parameters), it was taking a few days to complete, horrible. Ed explained everything to me, patiently, by email; also on magic multipliers for fast indexing; so most of the increased strength of the latest pattern-based Cake is thanks to him. I only just "finished" adding patterns and training, so there are still a lot of things that could be tried - but my gut feeling is that none of them will make such a large difference any more (things like adding more games, adding games of the latest engines = higher quality, adding games with more variety in openings, adding games of weaker engines... ).
http://laatste.info/bb3/viewtopic.php?f=53&t=6786
Code for 3-valued indices:
https://github.com/rhalbersma/scan/blob ... c/eval.cpp (my mirror of Fabien's Scan)
http://laatste.info/bb3/viewtopic.php?f ... 06#p112702
Some notes on eval types + a working Keras/Tensorflow implementation to tune Ed's 10x10 eval (30 mins on 240M positions, I still need to put this in a proper GitHub repo, no time lately)
http://laatste.info/bb3/viewtopic.php?f=53&t=8298
http://laatste.info/bb3/viewtopic.php?f=53&t=8327
